US 7058590 B2 Abstract A method and system for processing test data relevant to specific behavior of visitors of a network accessible site, such as a website, includes a number of components. A first component is configured to determine an initial estimate of visitor behavior on the basis of pre-testing information. Such information may be entered by a manager of the site. A second component is configured to generate updates of the estimate in response to monitored behavior. Bayesian estimation may be employed in this component. The third and fourth components cooperate to dynamically adjust a measure of the required test sample size of the visitors so as to maintain a target statistical confidence level. The third component utilizes systematic sampling, while the fourth component uses negative binomial sampling.
Claims(15) 1. A computerized method of processing test data that is relevant to specific behavior of visitors of at least one network accessible site comprising the steps of:
receiving pre-test information related to determining an estimation of anticipated behavior of said visitors to said at least one network accessible site;
monitoring actual behavior of said visitors upon accessing said at least one network accessible site;
employing said pre-test information and a Bayesian estimation approach to said monitoring of said actual behavior so as to provide updated estimations of subsequent visitor behavior, said behavior being related to whether said visitors are converted while accessing said at least one network accessible site, said pre-test information and said updated estimations being related to determination of conversion rates;
determining a required test sample size for said monitoring of said actual behavior in order to provide said updated estimations of subsequent visitor behavior, including adaptively adjusting said determination of said required test sample size on a basis of achieving a target confidence level regarding said updated estimations; and
determining a post-test estimation of said subsequent visitor behavior, including selecting between using a systematic sampling approach and using a negative binomial sampling approach, said negative binomial sampling approach being selected in response to unavailability of said required test sample size for said monitoring;
wherein an automatic process is defined by integration of said steps of receiving, monitoring and employing with said steps of determining a required test sample size and determining said post-test estimation, said post-test estimation being configured to provide a basis for determining interactions during subsequent visits to said at least one network accessible site.
2. The computerized method of
3. The computerized method of
4. The computerized method of
5. The computerized method of
6. The computerized method of
7. A computerized method of processing test data that is relevant to specific behavior of visitors of at least one network accessible site comprising the computer-executable process steps of:
detecting conversions at said at least one network accessible site, where said conversions are specific interactions that are entered by said visitors;
utilizing automated negative binomial sampling to determine an adaptive minimum sample size for estimating a conversion rate for subsequent visitors, such that an estimated conversion rate is available for basing determinations regarding subsequent operations of said at least one network accessible site, including basing said adaptive minimum sampling size and an estimate of said conversion rate on said detecting of said conversions and upon a measure of confidence; and
activating automated systematic sampling to determine said adaptive minimum sampling size for estimating said conversion rate upon detection of acquiring a threshold number of samples, said negative binomial sampling being disabled following said acquiring said threshold number of samples.
8. The computerized method of
9. The computerized method of
10. The computerized method of
11. A computer-usable medium containing computer-executable code for processing test data that is relevant to specific behavior of visitors of at least one network accessible site comprising:
a first module component for determining an initial conversion-related estimate on a basis of pre-testing information;
a second module component for generating updates of said conversion-related estimate in response to monitored behavior of said visitors of said at least one network accessible site, thereby enabling operations at said at least one network accessible site to be determined on a basis of said conversion-related estimate;
a third module component for dynamically adjusting a measure of a required test sample size of said visitors while maintaining a target statistical confidence level, said third module component applying systematic sampling; and
a fourth module component for utilizing negative binomial sampling to generate measures of said test sample size for occurrences in which said required test sample size of said third module component is unsatisfied.
12. The computer-readable medium of
13. The computer-readable medium of
14. The computer-readable medium of
15. The computer-readable medium of
Description The invention relates generally to processing test data that is relevant to a specific behavior of visitors at a network accessible site, such as a website available via the Internet, and more particularly to determining conversion rates of visitors to such sites. With the widespread deployment of the global communications network referred to as the Internet, the capability of providing electronic service (e-service) has become important to even well-established traditional business entities. An “e-service” is an on-line service that markets goods or services, solves problems, or completes tasks. E-services are accessible on the Internet by entering a particular Uniform Resource Locator (URL) into a navigation program. Operators of e-services are often interested in inducing visitors of a website to act in a certain manner. For example, an operator (i.e., e-marketer) may be interested in the sale of goods or services to visitors or may merely request that visitors register by providing selected information. When a visitor acts in the desired manner, the event may be considered (and will be defined herein) as a “conversion.” The ratio of visitors who are converted to the overall number of visitors is referred to as a “conversion rate.” Presently, conversion rates at Internet websites are relatively low, typically in the range of two percent to four percent. For various reasons, managers of websites are interested in accurate measures of conversion rates. For example, a change in a conversion rate may be used as a measure of the effectiveness of a promotion. Promotional offers are often presented to visitors in order to induce the visitors to interact with the website in a desired manner, e.g., register or purchase a product. Promotional offers include providing a discount on the price of the product being sold, providing free shipping and handling of the product, and/or providing a cost-free item. The typical goal of a promotion campaign plan is to increase the conversion rate in a cost-efficient manner. There are a number of considerations in determining estimations of conversion rate or other estimations of anticipated behavior by visitors to a network accessible site. On some occasions, there is available pre-testing information regarding the conversion rate of a website. There may be a relatively low or relatively high level of confidence in the accuracy of such information. Thus, one consideration is whether to incorporate the pre-testing information into the process of determining conversion rate. A second consideration is the selection of an approach for updating estimations. Yet another consideration involves selecting the sample size in testing visitors. Given the fact that each additional visitor that is tested causes a marketer to incur an additional cost and a potential loss in market opportunity, an important issue is determining how large the sample size needs to be in order to achieve a target level of confidence. A fourth consideration regards the methodology for sampling visitors for the testing. What is needed is a method and system which address these considerations in the estimations of anticipated visitor behavior. An adaptive testing approach utilizes at least some of four components that are cooperative in providing behavioral estimations that satisfy a required level of confidence of accuracy. As a first component of a system or method, the process is configured to determine an initial estimation on a basis of pre-testing information. For example, an e-marketer's prior knowledge may be incorporated into an initial conversion rate estimation by characterizing the knowledge with a suitable probability distribution. A second component is configured to generate updates of the estimation in response to monitored behavior of visitors to a network accessible site, such as a website. In one approach, the second component utilizes Bayesian estimation to provide updated estimations of subsequent visitor behavior. In third and fourth components, a minimum test sample size is determined while maintaining a target statistical confidence level. This determination is also adaptive, so that the measure of required test sample size is dynamically adjusted upwardly or downwardly in response to testing conditions. The third component uses systematic sampling. The fourth component is configured to utilize negative binomial sampling that is based on achieving the required confidence level. Within the first component, the prior knowledge of a manager of a site is entered and utilized. Ideally, the prior knowledge is an estimation of the visitor behavior (e.g., conversion rate), with the estimation being within a target confidence interval. However, in some situations, the manager is not able to provide an estimation of the conversion rate. Instead, another type of information may be available. For example, the manager may specify a conversion rate mean and a standard deviation, so that parameters of a prior distribution of the conversion rate can be determined using Bayes inference. In another possibility, the manager may specify a range of the conversion rate by a confidence interval. Again, Bayes inference may be used to determine the parameters of the prior distribution. After observations of visitor behavior are obtained, a Bayes estimator may be used to provide automatic updates of the estimation of the conversion rate or other behavioral parameter of interest. In one embodiment, the point estimation is an average of the pre-test estimation and a maximum likelihood estimate that is a result of the observed behavior. Bayes estimation is especially useful if there is prior knowledge and only a small sample of observations, since a small sampling is susceptible to inaccuracies. Regarding the determination of sampling size, the target number of successes (e.g., conversions) can be determined using systematic sampling at the third component. For example, from a probability criterion, a sample size may be identified as a ceiling. Then, from the expected number (N) of visitors, a requirement of the sampling pattern may be determined by dividing the expected number by the ceiling of the sampling size. A shortcoming of this systematic sampling approach is that there is a concern that the expected number of visitors will not be reached, so that the calculated test sample size will not be reached. In the fourth component, the shortcoming of the systematic sampling is addressed. Specifically, negative binomial sampling is utilized. The measure of the minimum test sample size therefore becomes dynamically adjustable by requiring the estimate of conversion rate to satisfy a particular statistical confidence level. The fourth module operates best in situations in which there may be a low number of visitors to a site. By integrating the four components, adaptive testing can intellectually and reliably address the main concerns of conversion estimation and testing. While the linkage of the first two components establishes the foundation for conversion rate estimation and updating, the linkage between the second and third components is a key to the dynamic sample size determination and allocation that provides managers with operational agility while maintaining targeted confidence. The linkage from the first and second components to the third and fourth components completes the automatic process in such a way that it provides seamless adaptive testing for predicting visitor behavior. With reference to The tool to be described below is intended to optimize the increased value derived from conversions of customers when promotions are offered to the customers. However, the adaptive testing invention may be used in other applications in which conversions are of significance to operators. A conversion is the act in which a visitor to a network site, such as a website, acts in a certain manner, such as purchasing a product or registering information. A campaign plan for determining which promotion should be presented to which customers is mathematically determined by an optimization engine The transactional database A promotions component A test marketing module Interaction with the design of a promotion campaign plan by a business manager takes place via a workstation The test results of an initial campaign model are passed from the test stage A number of actions take place within the campaign definition stage Within the testing stage The optimization stage As previously noted, the execution of the optimized plan utilizes the storefront A number of assumptions will be made in the description of the testing stage A “combination” will be defined herein as a segment-promotion pair. That is, each combination includes one customer segment that was defined in component Another assumption is that behaviors of visitors in the same combination are independent of each other and the individual conversion status (Y A. Incorporation of Prior Knowledge As previously described, the first component It is possible that the prior knowledge that is incorporated at step
Confidence intervals of the conversion rate θ may also be identified. A confidence interval of θ has a lower limit {circumflex over (θ)} As is known in the art of statistical economics, the confidence interval of confidence level 1−αis:
Rather than the normal approximation of the confidence interval, a more exact determination can be made. For the sample Y having a sample size of n and having y conversions, the upper and lower limits of the confidence interval can be determined by the following equations:
Thus far, the implementation of step
After the parameters are calculated, Bayesian estimation may be used to compute a point estimation. The techniques will be described in the following section, since the Bayesian estimation may also be used in the updates of the conversion rate calculation as testing is implemented. B. Bayesian Update Estimation Referring to If no prior knowledge is available, the parameters may be estimated using empirical Bayes analysis, which will not be described in detail in this document. The update of the estimate of conversion is represented by step C. Systematic Sampling Step Within the systematic sampling, a probability criterion is proposed and it is assumed that behaviors of visitors within each combination are independent of each other and that the conversion rates among combinations are not correlated. Regarding the probability criterion, for a given confidence level (1−α) and for an upper bound (d) of the distance between the estimate {circumflex over (θ)} If the total number of expected visitors in a particular combination is N, the systematic sampling scheme is to sample n* visitors from N total visitors. Taking d=floor [N/n*], the scheme is to generate a random start s from integers {1, 2, . . . , d}, and make offers to visitors s, s+d, s+2d, . . . , s+(n*−1)d. As previously noted, the shortcoming of the systematic sampling scheme is that the required sample size may only be reached if and when the N total visitors of a particular combination have visited the website. D. Negative Binomial Sampling It is supposed that the sequentially observed conversions Y Thus, it is important to determine the number of conversions m. The following probability criterion is used in the determination of the required success number m for a given precision level e and for a given confidence level 1−α: For the decomposition:
It should be noted that the confidence interval of θ in the probability criterion is not symmetrical about θ, since it is generally true that the absolute of ε Using the negative binomial sampling approach, inputs may include (1) the conversion number m, (2) the lower precision level ε After the total number m of conversions are detected during the testing stage E. Customer Allocation, Sequential Testing and Termination During the testing stage, there are number of considerations that must be addressed. One consideration is the technique for allocating promotions to arriving visitors within a particular customer segment. In some situations, the assumed conversion rates are not informative. For example, an e-marketer may not be able to provide any relevant information. In such situations, there are advantages to allocating all promotions to the arriving visitors alternately during the testing stage. Thus, if there are two promotions, the odd numbered visits result in presentation of the first promotion, while the even numbered visits trigger presentation of the second promotion. This achieves some randomization effect which can reduce unaware biases. On the other hand, if informative inputs on the assumed conversion rates are available for the different combinations, a proportional sampling scheme may be implemented in the allocation approach. For each combination of a customer segment and a promotion, an index (c) can be assigned on the basis of c=(i, k), where i is the customer segment and k is the promotion. Upon reaching a closing time (t), an attained confidence interval length D(c, t) is computed. Then, termination occurs for those c's in which the “convergence” of D(c, t) has been reached. One termination criterion for convergence is whether the variable moving average reaches a threshold. That is, termination occurs for those combinations that satisfy: Upon the termination of any combination, the resulting conversion rate is used to reevaluate the sample size requirement for additional sampling needs. In some occasions, the termination period will be reestablished. As a note, it may be beneficial to store all raw conversion data in a log for subsequent use. Patent Citations
Non-Patent Citations
Referenced by
Classifications
Legal Events
Rotate |